perf: four benchmarks remain 1.4-2.2x Node with no owning issue (pipeline, churn, tree_wide, cycles)

[proggeramlug]

The 2026-08-17 four-engine sweep leaves four rows where Perry is meaningfully behind Node and no issue owns them. interp/iso_miss are #8202 and fib40 is #8175; these are the remainder.

benchmark	Perry	Node	P/Node	Perry instr	Node instr
pipeline	0.2043	0.0950	2.15×	3.893 B	1.220 B
churn	0.3300	0.1666	1.98×	5.006 B	2.681 B
tree_wide	1.2965	0.9015	1.44×	26.808 B	20.262 B
cycles	0.1035	0.0736	1.41×	2.132 B	0.669 B

All four are instruction-bound rather than scheduling-bound — Perry executes 1.3–3.2× the instructions Node does — so the work is to emit less, not to schedule better.

Notes per row, from the sweep:

• pipeline is the worst uncovered gap. It was 3.3585 s at 08-14 and 0.2002 s at 08-15, so the catastrophic phase is long fixed (perf: pipeline's +3.95% at #8084 is unattributed — it is NOT the callee rooting (IR is byte-identical) #8243 closed); this is the steady-state residue. Perry beats scriptc here 0.2043 vs 0.4461, and Porffor cannot run it at all, so the gap is specifically against V8.
• churn also carries the perf(gc): #8277's object-start bitmap costs 8-10% wall on allocation-heavy workloads #8288 allocation-bitmap regression (+8.2% since 08-15). Fixing perf(gc): #8277's object-start bitmap costs 8-10% wall on allocation-heavy workloads #8288 should be sequenced first, then this re-measured — the two are additive and it would be easy to double-count.
• cycles has the largest instruction ratio of the four at 3.19× Node for 1.41× wall, which suggests the emitted work is cheap per instruction. Possibly the most tractable.
• tree_wide is the mildest ratio but the largest absolute time in the corpus at 1.30 s.

Not a regression report — all four improved or held since 08-15 (tree_wide −24.8%, cycles +1.4%, pipeline +2.0%, churn +8.2% from #8288). Filed so the remaining engine gap is tracked rather than implicit, now that the campaign has cleared the catastrophic rows.

Suggested order: #8288 first (it contaminates churn), then cycles as the cheapest probe into where the extra instructions come from, then pipeline.

Sweep evidence: gc-handoff/current-sweep-2026-08-17/ (RESULTS.md, results-clean.json, report.html).

[proggeramlug]

Progress: #8301 has merged. On the checked-in cycles reproducer, memoizing typed layouts by immutable ShapeId reduced the five-run median from 2.177B to 1.804B retired instructions (-17.1%) with unchanged output.

Keeping this umbrella issue open: pipeline, churn, and tree_wide remain, and cycles still needs a clean four-engine re-sweep to establish its new wall-time and Node ratios.